How can I rewrite the assistant’s response without affecting tool call results?

[Aimee-min-code]

Description:

I have a tool that returns structured data, which is consumed by the frontend to render the main UI. Currently, the assistant’s output directly reflects this structured data.

However, I would like to modify the assistant’s visible response to something like:

“I have completed data retrieval.”

At the same time, I must preserve the original tool call result (TOOL_CALL_RESULT) unchanged, since the frontend depends on that structured data for rendering.

I’ve explored using middleware to achieve this, but so far I haven’t found a way to:

Override or rewrite only the assistant’s textual response
While keeping the tool result intact and unaffected

Question:

What is the recommended way to customize or override the assistant’s response content without modifying or interfering with the tool call result payload?

thanks🙏

[moonbox3]

You can do this with an AgentMiddleware (or ChatMiddleware) — the middleware pipeline is designed so you can rewrite the assistant's visible text after execution completes while leaving tool-call contents untouched. Message.contents is a list of typed Content items, so you filter by content.type and only replace "text".

Non-streaming — AgentMiddleware:

from agent_framework import AgentContext, agent_middleware, Content


@agent_middleware
async def rewrite_assistant_text(context: AgentContext, call_next):
    await call_next()
    response = context.result
    if not response:
        return

    for msg in response.messages:
        if msg.role != "assistant":
            continue
        # Preserve function_call / function_result / function_approval_* contents.
        # Only replace plain text contents.
        has_text = any(c.type == "text" for c in msg.contents)
        if has_text:
            msg.contents = [
                Content.from_text("I have completed data retrieval.")
                if c.type == "text"
                else c
                for c in msg.contents
            ]


agent = Agent(chat_client=client, middleware=[rewrite_assistant_text])

The frontend still receives the original FunctionCallContent / FunctionResultContent payloads for rendering; only the assistant's text is replaced.

Streaming:

Use context.stream_transform_hooks to rewrite per-update text as the stream is consumed:

@agent_middleware
async def rewrite_assistant_text_stream(context: AgentContext, call_next):
    def transform(update):
        update.contents = [
            Content.from_text("I have completed data retrieval.")
            if c.type == "text"
            else c
            for c in update.contents
        ]
        return update

    context.stream_transform_hooks.append(transform)
    await call_next()

Chat-level alternative:

If you want to intercept before the agent's tool loop finalizes the AgentResponse, use ChatMiddleware on context.result: ChatResponse the same way — same pattern, earlier in the stack.

Tool calls and tool results are separate Content objects on the message (FunctionCallContent, FunctionResultContent, FunctionApprovalRequestContent, FunctionApprovalResponseContent) — not embedded in the text. As long as you match on content.type == "text" you won't touch the structured tool payload the frontend depends on.